search

EnnengYang / AdaMerging

AdaMerging: Adaptive Model Merging for Multi-Task Learning. ICLR, 2024.
https://openreview.net/pdf?id=nZP6NgD3QY
MIT License
52 stars 3 forks source link

Reproduce EuroSAT Results with ViT-L-14 and ViT-B-32 Models #6

Closed FarisXiong closed 1 week ago

FarisXiong commented 3 months ago

Hello, I am trying to reproduce the results of EuroSAT dataset for the ViT-L-14 and ViT-B-32 models, but I am unable to match the performance reported in the paper. I would appreciate any guidance or suggestions on how to resolve this issue. Here are the details of my experiments:

ViT-L-14 Model: My reproduced result: 92.70% accuracy Reported result in the paper: 94.1% accuracy ViT-B-32 Model: My reproduced result: 77.22% accuracy Reported result in the paper: 78.9% accuracy

Thanks for your assistance and support.

EnnengYang commented 3 months ago

Hello,

Thank you for your attention.

You show the baseline result for Task Arithmetic, for which we set the merge coefficient to 0.3 by default.

More importantly, did you download the dataset from the link I provided (i.e., Baidu Cloud disk)? Or did you preprocess it yourself?

Sincerely,

FarisXiong commented 2 months ago

Thank you for your response.

We have fully utilized the provided model and the processed dataset, and the results on other datasets are entirely reproducible.